From the NSA Cloud to LLM Guardrails

Top-Line Summary

David Bauer, who got his AWS account directly from Jeff Bezos before the platform launched publicly, joined the show to walk through his path from running a BBS in the 1980s to building the NSA’s first cloud compute environment in tractor trailers. The conversation landed on his current work: embedding LLMs inside MCP servers to create decision intelligence layers with audit trails, federated learning across data silos using homomorphic encryption, and the still-unsolved problem of getting AI agents to reliably follow instructions.

Show Video

The “Pre-Show” Context

Brett confessed that Claude Code has fully replaced his Battlefield Six addiction — though he’s now more sleep-deprived than when he was gaming. David asked if Claude was playing the game for him, which would’ve been a much better story. Travers predicted a new clinical syndrome for developer addiction is coming.

The Engineering Rundown

• The VIC-20-to-BBS Origin Story (01:30) David has been programming since 1982, starting on a VIC-20 his mom brought home from the IRS. He ran Texas Outlaws BBS through the eighties - a piracy board where you had to call someone and ask them to swap the floppy disk. Brett one-upped with his Commodore 64 tape drive days, typing programs out of Byte Magazine, and the trick of punching a notch in 5.25" floppies to make them double-sided. David still has his official floppy hole puncher. Brett still has the original Commodore hardware in his basement.

• Building the NSA’s Cloud in Tractor Trailers (06:00) Before “cloud” meant what it means today, David was working under General Alexander at the NSA to build the cloud compute environment - which in 2005 was really a big data platform focused on getting petabytes of data into portable optimized data centers (tractor trailer bodies with 50U racks). They flew one on an Antonov to Afghanistan, where David went with the system. The team had spent millions on 10-gigabit top-of-rack networking, but the base’s overall throughput was about one gigabit for every computer on site. Bandwidth was, and apparently still is, the real bottleneck. Also: don’t leave the trailer door open in Afghanistan - it snows.

• Getting an AWS Account from Jeff Bezos (08:30) During a Silicon Valley tour in 2005, David’s team met Jeff Bezos and the Amazon crew. Bezos asked if he wanted to try it, turned to an engineer, and said “give him an account.” David still has that same account - he just doesn’t use it because he doesn’t want to pay for it. He wishes there was a “member since” badge. Brett admitted he needs to go look up his own account creation date.

• Parallel Discrete Event Simulation at 10 Million CPU Cores (13:00) David’s PhD work produced what may be the world’s largest monolithic application - a parallel discrete event simulator that ran on 9.9 million CPU cores at Lawrence Livermore. A sequential version takes about 10 lines of code; the parallel version is 50,000 lines. The origin story: his advisor left for Germany for two weeks, David and a fellow grad student house-sat the advisor’s terrible dog, couldn’t make the existing software work, and clean-room rewrote it from scratch in those two weeks. It took two more years to stabilize. Use cases ranged from chip design (now replaced by AI) to simulating billions of cell phones with actual partial wave dynamics.

• COVID Pandemic Modeling with Human Behavior (15:30) David’s team applied their simulation platform to COVID by discretizing the SEIR differential equations and layering in human behaviors - mask compliance, vaccination likelihood, whether someone could actually work from home. At peak they were modeling about 5-6 billion people at the individual level, tracking movement from room to room, car to bus to grocery store. US Northern Command and the National Guard used it to plan PPE and medical staff distribution. The Army used it to predict what would happen when 2,000 soldiers went to bars in Poland during a training exercise. The intelligence community used it to assess force readiness in Ukraine, Belarus, and Russia - David has a laptop with “about 400 viruses” from collecting Russian state agency data.

• Decision Intelligence and LLMs Inside MCP Servers (18:00) David’s current work flips the typical MCP model. Instead of just exposing tools to an external LLM, his team puts LLMs inside the MCP server itself. These internal models track the organization’s decision-making process, creating a dataset about how decisions are actually made - where bottlenecks are, where decisions go to die, how static or plastic the processes are. The system requires sources of evidence, written records of why a decision matters, and who approves or rejects it. Many organizations reveal their maturity level immediately by saying “that means you’re gonna connect to my email” — meaning their entire decision process is someone emailing yes or no.

• Federated Learning with Homomorphic Encryption (24:30) Rather than centralizing data (“you can put it all on Snowflake and wait for it to get hacked,” David slid in), the team leaves data where it is and trains locally. Every 10-20 epochs, model weights and losses are exchanged with an arbiter that performs federated averaging. Since the threat surface shifts from data to models - leaked weights can be used to reconstruct training data, which is how some cheaper Chinese lab models get built - they protect everything with homomorphic encryption (encrypted over the wire, on disk, and in memory), garbled circuits to fuzzy models against reconstruction, and differential privacy.

• The Scaffolding Problem and Getting LLMs to Follow Rules (30:00) Brett shared his frustration with AI agents that follow instructions three out of five times and then ask permission on the fourth. David compared it to having an extra child. The consensus: the company with the best scaffolding wins. David’s team constantly discovers new corner cases in decision-making processes - they thought they’d just feed DOD doctrine into an LLM and be done in two days, but humans are remarkably good at working around policy. The data poisoning problem is real too: if you’re learning from bad decision-makers, you just learn bad habits. They use the LLMs to classify whether out-of-compliance decisions are genuinely bad or just unconventional, creating clean training data for fine-tuning.

• Mamba 3 and the Open Source AI Community (44:00) David flagged the Mamba 3 model release under Apache license, focused on speeding inference on smaller hardware. The open source community is still doing interesting work, even as the hyperscalers pour hundreds of billions into infrastructure. Brett noted AWS/Kiro just added Qwen, Minimax, and other models to Bedrock - a surprise given their massive Anthropic investment, but classic AWS diversification.

Off-the-Clock Recommendations

• We Are Legion (We Are Bob) by Dennis E. Taylor — Brett’s on a binge of the Bobiverse series. Light sci-fi, contributing to sleep deprivation.
• Vibe Coding by Steve Yegge and Gene Kim — Not what you think from the title. Brett compared it to Gene Kim’s Phoenix Project. Another page-turner.
• Project Hail Mary (2026 film) — Travers is going to see it opening Friday. Brett loved the book and is hoping the movie does it justice. IMAX was discussed.
• The Premonition by Michael Lewis — Travers is reading it, which is why he had pandemics on the brain during the simulation discussion.
• Go climb a mountain — David’s first recommendation. Get away from screens and see things more than a couple feet from your face.
• Crochet — David’s second recommendation, specifically for carpal tunnel prevention. The blanket on his couch is proof. Brett got beginner crochet kits for Christmas from a tech Discord where 50%+ of the members crochet or knit.